AI terms

The AI vocabulary changes every six months. This page pins down the words this Playbook uses so the rest of it reads consistently. If you're new to AI tooling, read this once before you start running skills.

Skill

A skill is a packaged prompt that does a specific job. Think Excel template, Word macro, or a saved query — someone built it once, you load it, run it on your data, get the work back.

Every entry under the Skills sidebar folder is a skill. Each one is one prompt, one job, one expected output.

Both runtimes now have a formal Skills feature (as of late 2025). In Claude Desktop, you create them in Settings → Features → Skills, or save them inside a Project. In OpenAI Codex, they live under the Skills tab in the left sidebar. Both apps load skills the same way: paste the skill's prompt body into a Skill, save, then run threads against it.

Prompt

A prompt is the actual text you paste into the AI. It tells the AI what role to play, what data to expect, what to produce. Every skill page in this Playbook has its prompt body in a copy-pasteable block — no <placeholders> to fill in, no "your data here" filler.

You can run a prompt as a one-off (paste it into a fresh chat, attach the file, read the output) or save it as a reusable Skill so you don't paste it every time. Save it.

Attachment

Attachments are the files you give the AI alongside the prompt. A CSV export from your ERP. A PDF spec sheet from a supplier. Last quarter's QBR deck. Every skill page in this Playbook lists exactly what to attach under Inputs to attach.

The AI reads attachments end-to-end — it doesn't skim. If you attach a 40-page PDF, every page gets read.

Context window

The context window is how much text the AI can hold in working memory at once. Current models — Claude Sonnet 4.6 and GPT-5.5 — sit around 200,000+ tokens, roughly a 500-page book.

You don't have to think about this for most skills. Where a skill needs more (e.g., a full year of CRM history), the skill page calls it out and tells you how to chunk the input.

Agent task

An agent task is a skill that takes multiple steps on its own before returning. Instead of "summarize this PDF," it's "read this PDF, then check the public-facing pricing page, then write me a one-page brief with both sources cited."

Some skills in this Playbook are framed as agent tasks — they call out to the web, read multiple files in sequence, or run a check-then-decide flow. You don't set anything up differently to run one; both Claude Desktop and Codex handle this automatically when the prompt asks for it.

Scheduled task

A scheduled task runs a skill on a recurring cadence without you starting it. Monthly distributor scorecards on the first Friday. Weekly competitive briefings every Monday at 8am.

• Claude Desktop schedules sit in the Scheduled tab in Claude Cowork. Cadences: hourly, daily, weekdays, weekly, manual. Important caveat: scheduled tasks only run while your computer is awake and the Claude Desktop app is open. Skipped runs re-execute when the app reopens.
• OpenAI Codex uses Automations, set per Skill. Cadences: hourly, daily, weekly, monthly. Codex automations behave the same way — they need the app running.

If you need a schedule that fires when your laptop is closed, see Routine below.

Routine (cloud-hosted schedule)

A routine is a scheduled task that runs on the AI vendor's servers instead of your laptop. Anthropic launched Cloud Routines in research preview in April 2026 — they package a prompt and inputs into an automation that runs on Anthropic infrastructure, on a schedule, an HTTP webhook, or a GitHub event. Daily run limits apply (Pro 5/day, Max 15/day, Team/Enterprise 25/day).

Cloud Routines are currently aimed at Claude Code (the developer CLI), not the marketing analyst flow. Most readers of this Playbook will use the desktop-based scheduled tasks above. If you grow into a workload where laptop-on-and-app-open is a real constraint, ask Jason.

Runtime

A runtime is the application you've installed on your computer. The two runtimes this Playbook supports:

• Claude Desktop (with Claude Cowork as the agentic mode inside it). Runs Claude Sonnet 4.6 and the Opus tier on Pro+. The default for editorial, analytical, and long-context work.
• OpenAI Codex. Runs GPT-5.5 (and GPT-5.5 Pro on Pro). Stronger pull on developer-style tasks (file system manipulation, code, structured automations).

Each skill page notes the recommended runtime. Most skills work in both. You don't have to pick before you start — install one, run a couple of skills, install the other if you hit a limitation.

Project (Claude) vs Skill (Codex)

A naming collision worth flagging:

• In Claude Desktop, the workspace where a saved skill lives is called a Project. The skill's prompt body goes into the Project's Set project instructions field. Files Claude should always have in mind go into Project knowledge.
• In OpenAI Codex, the workspace is called a Skill directly. The prompt body goes into the Instructions field.

Same idea, different label. This Playbook uses skill as the universal unit. Where the install or schedule steps differ between runtimes, we call them out.

What we don't use

You'll see plenty of marketing copy that uses these terms — this Playbook avoids them on purpose:

• "AI assistant" / "AI copilot" — too vague. We name the runtime (Claude Desktop, Codex) and the skill (the specific prompt) so it's clear what's running.
• "Magic" / "transform" / "revolutionize" — disqualifiers. If a vendor pitches you with these, ask them what specifically the AI does and what specifically it doesn't.
• "Agentic" as a marketing word — we use agent task for the specific shape (multi-step, calls out to other systems) and avoid the vibes word.

Where to go next

• Installing Claude Desktop — set up the Anthropic runtime in about 10 minutes.
• Installing OpenAI Codex — set up the OpenAI runtime in about 10 minutes.
• Using a skill — the four-step pattern for running any skill in this Playbook.